Fire hazard level evaluation method and system, storage medium and equipment

By adopting a continuous learning method in the assessment of fire hazard level, dividing the training data subsets and updating the model with distillation loss and task loss, the 'catastrophic forgetting' problem in the face of changing conditions is solved, and the fire prediction is improved Accuracy and stability.

CN118965141APending Publication Date: 2024-11-15YUNNAN POWER GRID CO LTD ELECTRIC POWER RES INST
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411001097.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-25
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

Traditional fire hazard level prediction methods are prone to ‘catastrophic forgetting’ when facing changing climates and vegetation states, and cannot effectively adapt to new conditions, resulting in a decrease in prediction accuracy.

Method used

Using a continuous learning-based method, the historical fire data set is divided into multiple training data subsets, and the neural network model is trained through different structural parameters and training strategies. The model weight is updated using a combination of distillation loss and task loss to ensure the model Keep knowledge of the old tasks while learning new tasks.

Benefits of technology

It improves the accuracy and efficiency of fire hazard level assessment, reduces the probability of the occurrence of ‘catastrophic forgetting’ phenomenon, and provides more reliable fire prevention and control support.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118965141A_ABST
    Figure CN118965141A_ABST
Patent Text Reader

Abstract

The embodiment of the invention discloses a fire hazard level evaluation method, which comprises the following steps: acquiring a historical fire hazard data set which comprises combustible material data, power transmission line data and corresponding fire hazard levels; dividing the historical fire data set into a plurality of training data subsets including a first task data subset and a second task data subset, the first task data subset and the second task data subset respectively use a first structure parameter and a first training strategy as well as a second structure parameter and a second training strategy to train an original neural network model; continuously learning the original neural network model according to the plurality of training data subsets to obtain a continuously learned neural network model; and determining fire hazard levels corresponding to the first task data subset and the second task data subset according to the continuously learned neural network model. According to the method, multi-task learning and continuous learning are combined, and the accuracy of fire hazard level evaluation is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of risk assessment, and in particular to a fire hazard level assessment method, system, storage medium and equipment. Background Art

[0002] In recent years, affected by climatic conditions and the custom of using fire in industrial and agricultural production, wildfires have frequently occurred near transmission lines, posing a serious threat to the safe operation of transmission lines. In particular, in recent years, many UHV transmission lines have tripped or been locked due to wildfires, and the situation of wildfire prevention and control on transmission lines is severe. With the rapid development of artificial intelligence technology, its application in various fields is becoming more and more extensive, among which fire prediction is also an important application scenario.

[0003] Traditional fire danger level prediction methods usually rely on static, predefined feature sets and models. These methods learn the laws of fire occurrence under specific conditions during the training process. However, when new monitoring data is continuously input and the conditions and patterns of fire occurrence change, traditional methods are often unable to adapt to such changes, resulting in a decrease in prediction accuracy. For example, the model learns to identify the relationship between specific types of vegetation dryness and wildfires in spring, but in autumn, as the climate and vegetation conditions change, the model may not be able to correctly identify the connection between new climate features and fires, and may have forgotten the relationship between vegetation dryness and wildfires learned in spring. Therefore, this application proposes a fire danger level assessment method based on continuous learning. Summary of the invention

[0004] Based on this, it is necessary to propose a fire hazard level assessment method to address the above issues.

[0005] A method for evaluating fire hazard level, comprising the following steps:

[0006] Acquire a historical fire data set, wherein the historical fire data set includes combustible material data, power transmission line data, and corresponding fire hazard levels;

[0007] Dividing the historical fire data set into a plurality of training data subsets, the plurality of training subsets comprising a first task data subset and a second task data subset, the first task data subset and the second task data subset respectively using a first structural parameter, a first training strategy and a second structural parameter, a second training strategy to train an original neural network model;

[0008] Continuously learning the original neural network model according to the multiple training data subsets to obtain a neural network model after continuous learning;

[0009] The fire hazard levels corresponding to the first task data subset and the second task data subset are determined according to the neural network model after continuous learning.

[0010] In the above solution, the historical fire data set is divided into a plurality of training data subsets, specifically including:

[0011] The historical fire data set is divided according to a time dimension or a space dimension to obtain a plurality of training data subsets.

[0012] In the above scheme, the continuous learning of the original neural network model according to the multiple training data subsets to obtain the neural network model after continuous learning specifically includes:

[0013] The original neural network model determines, based on the sample data in the second task data subset, a first fire hazard level output by the sample data in the first task;

[0014] The neural network model determines, based on the sample data in the second task data subset, a second fire hazard level output by the sample data in the first task and a third fire hazard level output in the second task;

[0015] Determine a distillation loss between the fire hazard level output by the original neural network model for the sample data in the first task and the fire hazard level output by the neural network model for the sample data in the first task;

[0016] Determining a task loss between a fire hazard level output by the sample data in a second task and a true label corresponding to the sample data;

[0017] According to the total loss value composed of the distillation loss and the task loss, the weight parameters of the neural network model are updated.

[0018] In the above scheme, the first fire hazard level and the second fire hazard level are determined according to the following formula:

[0019] y o =SV n (X new ,θ shared ,θ old )

[0020] y n =SV n (X new ,θ shared ,θ old )

[0021] Among them, y o is the fire hazard level output by the sample data in the first task of the original neural network model, y n is the fire hazard level output by the sample data in the second task of the neural network model, SV nis the neural network model, θ old is the weight parameter corresponding to the first task, θ shared is the shared parameter between the second task and the first task.

[0022] In the above scheme, the total loss value formed by the distillation loss and the task loss and updating the weight parameters of the neural network model specifically include:

[0023]

[0024] in, are the shared parameters of the neural network model between the second task and the first task, is the weight parameter of the neural network model corresponding to the first task, is the weight parameter of the neural network model corresponding to the second task, λ o To set the hyperparameters of the balanced distillation loss and task loss, is the distillation loss, is the task loss between the fire hazard level output by the sample data on the second task and the true label corresponding to the sample data, is the regularization term, θ s To update the shared parameters of the neural network model between different second tasks and the first task, θ o is the weight parameter of the neural network model corresponding to the first task before updating, θ n It is the weight parameter corresponding to the second task of the neural network model before updating.

[0025] In the above solution, determining the fire hazard levels corresponding to the first task data subset and the second task data subset according to the neural network model after continuous learning specifically includes:

[0026] The fire hazard levels corresponding to the first task data subset and the second task data subset are evaluated based on the comparison relationship between the fire occurrence probability output by the neural network model and the preset threshold, and the fire hazard levels include four levels: a first hazard level, a second hazard level, a third hazard level and a fourth hazard level.

[0027] In the above scheme, the fire hazard level corresponding to the first task data subset and the second task data subset is evaluated according to the comparison relationship between the fire occurrence probability output by the neural network model and the preset threshold, specifically including:

[0028] When the fire occurrence probability output by the neural network model is in the range of 0.9-1, the fire hazard level is the first hazard level;

[0029] When the fire occurrence probability output by the neural network model is in the range of 0.7-0.9, the fire hazard level is the second hazard level;

[0030] When the fire occurrence probability output by the neural network model is in the range of 0.6-0.7, the fire hazard level is the third hazard level;

[0031] When the fire occurrence probability output by the neural network model is in the range of 0-0.6, the fire hazard level is the fourth hazard level.

[0032] The present application also proposes a fire hazard level assessment system, the system comprising: a data acquisition unit, a data classification unit, a model training unit and an assessment unit;

[0033] The data acquisition unit is used to acquire a historical fire data set, wherein the historical fire data set includes combustible material data, power transmission line data and corresponding fire hazard levels;

[0034] The data classification unit is used to divide the historical fire data set into a plurality of training data subsets, wherein the plurality of training subsets include a first task data subset and a second task data subset;

[0035] The model training unit is used to continuously learn the original neural network model and the neural network model being learned on the second task according to the multiple training data subsets to obtain the neural network model after continuous learning;

[0036] The evaluation unit is used to determine the fire hazard level corresponding to the second task data subset according to the neural network model after continuous learning.

[0037] The present application also proposes a readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the following steps:

[0038] Acquire a historical fire data set, wherein the historical fire data set includes combustible material data, power transmission line data, and corresponding fire hazard levels;

[0039] Dividing the historical fire data set into a plurality of training data subsets, the plurality of training subsets comprising a first task data subset and a second task data subset, the first task data subset and the second task data subset respectively using a first structural parameter, a first training strategy and a second structural parameter, a second training strategy to train an original neural network model;

[0040] Continuously learning the original neural network model according to the multiple training data subsets to obtain a neural network model after continuous learning;

[0041] The fire hazard levels corresponding to the first task data subset and the second task data subset are determined according to the neural network model after continuous learning.

[0042] The present application also proposes a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the following steps of the computer program:

[0043] Acquire a historical fire data set, wherein the historical fire data set includes combustible material data, power transmission line data, and corresponding fire hazard levels;

[0044] Dividing the historical fire data set into a plurality of training data subsets, the plurality of training subsets comprising a first task data subset and a second task data subset, the first task data subset and the second task data subset respectively using a first structural parameter, a first training strategy and a second structural parameter, a second training strategy to train an original neural network model;

[0045] Continuously learning the original neural network model according to the multiple training data subsets to obtain a neural network model after continuous learning;

[0046] The fire hazard levels corresponding to the first task data subset and the second task data subset are determined according to the neural network model after continuous learning.

[0047] The embodiment of the present invention has the following beneficial effects: firstly, a historical fire data set is obtained, wherein the historical fire data set includes combustible data, power transmission line data and corresponding fire hazard levels; the historical fire data set is divided into multiple training data subsets, wherein the multiple training subsets include a first task data subset and a second task data subset, wherein the first task data subset and the second task data subset respectively use a first structural parameter, a first training strategy, a second structural parameter and a second training strategy to train an original neural network model; the original neural network model is continuously learned according to the multiple training data subsets to obtain a neural network model after continuous learning; and the fire hazard levels corresponding to the first task data subset and the second task data subset are determined according to the neural network model after continuous learning. The present invention combines historical fire data sets and multi-task learning, and can improve the accuracy and efficiency of fire hazard level assessment by dividing training data subsets and applying different structures, training strategies and continuous learning techniques. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0049] in:

[0050] Figure 1 A schematic diagram of a method for evaluating a fire hazard level in one embodiment;

[0051] Figure 2 A schematic diagram of a process for processing a historical fire data set in one embodiment;

[0052] Figure 3 A schematic diagram of a process of dividing a historical fire data set in one embodiment;

[0053] Figure 4 A schematic diagram of a deep neural network model structure in one embodiment;

[0054] Figure 5 Based on Figure 4 Flow chart of the fire hazard level assessment method. DETAILED DESCRIPTION

[0055] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0056] In the following description, a large number of specific details are given in order to provide a more thorough understanding of the present invention; however, it is obvious to those skilled in the art that the present invention can be implemented without one or more of these details; in other examples, in order to avoid confusion with the present invention, some technical features known in the art are not described, and it should be understood that the present invention can be implemented in different forms and should not be interpreted as limited to the embodiments presented here; on the contrary, providing these embodiments will make the disclosure thorough and complete and fully convey the scope of the present invention to those skilled in the art.

[0057] The purpose of the terms used herein is only to describe specific embodiments and is not intended to be limiting of the present invention. When used herein, the singular forms "a", "an" and "said / the" are also intended to include the plural forms, unless the context clearly indicates otherwise, and it should also be understood that the terms "consisting of" and / or "comprising", when used in this specification, determine the presence of the features, integers, steps, operations, elements and / or parts, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, parts and / or groups. When used herein, the term "and / or" includes any and all combinations of the relevant listed items.

[0058] Traditional machine learning methods face a major challenge in predicting fire danger levels - catastrophic forgetting. Catastrophic forgetting is a common problem in the fields of machine learning and deep learning. When a model is trained to handle a second task, it often loses its memory of the previous learning task, causing the model's performance on the first task to drop sharply. This problem is particularly prominent in the field of fire prediction, because the occurrence of fire is affected by many factors, including but not limited to climate conditions, vegetation status, human factors, etc., and these factors change with time and region. Traditional machine learning methods find it difficult to maintain long-term memory of all relevant features when processing such continuous and changing data, and are prone to "catastrophic forgetting".

[0059] In order to solve this problem, this application proposes the application of continuous learning method in fire prediction. Continuous learning aims to enable the model to retain knowledge of the first task while learning the second task, thereby overcoming the problem of catastrophic forgetting. By introducing new learning strategies and model architectures, the continuous learning method is expected to improve the accuracy and stability of fire hazard level prediction and provide more reliable technical support for fire prevention and control.

[0060] Before introducing the continuous learning method provided in the embodiment of the present application, some concepts involved in the embodiment of the present application are first introduced:

[0061] Continuous learning usually refers to allowing machine learning models to continuously learn new information over time while retaining and improving their understanding of old knowledge. For example, after a machine learning model learns to identify the relationship between the dryness of a specific type of vegetation and wildfires in the spring, it learns the connection between new climate features and fires in the fall. At this time, the machine learning model will not forget the relationship learned in the spring.

[0062] The following is an example of an application scenario where this method is applicable: for the forest fire hazard coefficient of the transmission line, or to evaluate and predict the fire hazard level in a low pressure environment, it is understandable that the factors of forest fire risk assessment related to the transmission line are complex, such as: human factors, meteorological factors, terrain factors, vegetation factors, flame combustion factors, etc., and for example: slope, slope aspect, altitude, combustible load, ignition point, calorific value, moisture content, spatial correlation, etc. The factor fields contained in the training data collected each time may be different, for example: the impact of different seasons on vegetation combustible related parameters, the difference in vegetation characteristics in different regions, and the difference in electricity consumption of different transmission lines will affect the prediction effect of the fire hazard level. In this application scenario, multiple training data subsets in the above-mentioned continuous learning method can be used to continuously learn the original neural network trained on the old task data subset and the neural network model that is learning on the new task, so that the neural network model can learn the common effective feature representations in the new task and the old task, that is, the neural network model can continuously learn new tasks while retaining the learning results of the old tasks.

[0063] In order to thoroughly understand the present invention, a detailed structure will be presented in the following description to illustrate the technical solution proposed by the present invention; optional embodiments of the present invention are described in detail as follows, but in addition to these detailed descriptions, the present invention may also have other implementation methods.

[0064] like Figure 1 As shown, in one embodiment, a fire hazard level assessment method is provided, and the fire hazard level assessment method includes steps S101 to S104, which are described in detail as follows:

[0065] S101, obtaining a historical fire data set, where the historical fire data set includes combustible material data, power transmission line data, and corresponding fire hazard levels;

[0066] Specifically, the above historical fire data set can be collected in different scenarios, such as: power transmission line forest scenario, wildfire scenario in low pressure environment, or power transmission line forest scenario in low pressure environment, etc. It is understandable that most overhead power transmission lines pass through forests or mountains with high surface vegetation coverage and harsh terrain conditions. During the high incidence of outdoor fires such as spring plowing, Qingming Festival, autumn harvest and Spring Festival, large-scale wildfires are prone to occur around overhead power transmission lines, thus affecting the life of the transmission lines.

[0067] Preferably, Figure 2The figure shows a schematic diagram of the processing process of the historical fire data set. The above-mentioned historical fire data set may include combustible data, transmission line data and fire hazard level; wherein the combustible data and transmission line data are training data, and the fire hazard level is a training label. The combustible data include, for example, vegetation type (expressed as x1), vegetation height (expressed as x2), combustion equivalent (expressed as x3), moisture content of combustibles (expressed as x4), ignition point (expressed as x5), calorific value (expressed as x6) and heat release rate (expressed as x7), etc. The transmission line data include, for example, voltage (expressed as x8), resistance (expressed as x9), reactance (expressed as x10), susceptance (expressed as x11), conductance (expressed as x12), etc.

[0068] Among them, the above-mentioned moisture content of combustibles refers to the percentage of moisture content of forest combustibles to dry weight, which can be obtained by the formula IR = (fresh weight of sample - dry weight of sample) / dry weight of sample × 100%. The flash point refers to the minimum temperature required for the surface of a material to ignite and continue to burn for a certain period of time under specified test conditions using an external heat source. The calorific value refers to the heat released by the complete combustion of 1 kilogram (cubic meter) of a certain solid (gas) fuel, which is called the calorific value of the fuel. The heat release rate can be obtained by testing with a cone calorimeter. The heat release rate (Heat Release Rate, HRR) is an important performance parameter that characterizes the intensity of fire, and its unit is kW / m2; the maximum value of HRR is the peak value of the heat release rate. The above-mentioned voltage is a potential difference or electric potential difference, which is a physical quantity that measures the energy difference caused by the different electric potentials of unit charges in the electrostatic field. Resistance can be the resistance per unit length of non-ferrous metal wires (including aluminum wires, steel-core aluminum wires and copper wires). The reactance of the power line is formed when a three-phase symmetrical alternating current passes through the wire, generating an alternating magnetic field around the wire. The conductivity of the overhead transmission line is mainly related to the line corona loss and the leakage resistance of the insulator. Susceptance is a parameter of a passive linear element in circuit theory, which is used to describe the circuit's response ability to alternating current, especially focusing on the influence of pure capacitive or inductive elements.

[0069] The above-mentioned combustible data and transmission line data can be obtained separately. For example, the combustible data can be collected according to different time resolutions such as hourly or daily, and the combustible data and transmission line data can be manually identified. Of course, the combustible data and transmission line data can also be packaged into a historical fire data set and obtained together. This historical fire data set can be expressed as The above-mentioned historical fire data set can be obtained from a database or file system, and the file system can include: FAT, NTFS, HFS and ext series, etc. The above-mentioned database can be a special database, for example: a more accurate combustible database for the research area established based on laboratory measurement methods. Of course, the database can also be a general relational database, such as: MySql, PostgreSQL, Oracle and SQLSever, etc.

[0070] S102, dividing the historical fire data set into a plurality of training data subsets, the plurality of training subsets including a first task data subset and a second task data subset, the first task data subset and the second task data subset respectively using a first structural parameter, a first training strategy and a second structural parameter, a second training strategy to train an original neural network model;

[0071] Specifically, the first task data subset is the old task data subset, and the second task data subset is the new task data subset. The collection time of the old task data subset is earlier than the collection time of the new task data subset. By dividing the historical fire data set into the old task data subset and the new task data subset according to the time dimension, the neural network model dynamically adapts to the changes in the time dimension, and is thus trained to adapt to new tasks in the time dimension, which effectively reduces the probability of occurrence of the "catastrophic forgetting" phenomenon, thereby improving the accuracy of fire hazard level prediction.

[0072] like Figure 3 As shown, the historical fire data set can be divided into multiple training data subsets by using the spatial dimension division method. Of course, the historical fire data set can also be divided into multiple training data subsets by using the time dimension division method. The multiple training data subsets here can be represented as TR i (i=1,2,...,m), and the neural network model after learning with the first training data set TR1 can be expressed as SV1, and the neural network model after learning with the second training data set TR2 can be expressed as SV2, and so on. m The learned neural network model can be expressed as SV m Therefore, the above-mentioned continuous learning method can be oriented to different spatiotemporal regions, that is, the continuous learning method can be extended to regions of different spatiotemporal scales, so that the neural network model can be continuously learned and updated in different spatiotemporal regions.

[0073] In some embodiments, the historical fire data set is divided into a plurality of training data subsets, specifically including:

[0074] The historical fire data set is divided according to the time dimension or the space dimension to obtain multiple training data subsets.

[0075] Preferably, the geographic parameter values ​​of the new task data subset may be different from the geographic parameter values ​​of the old task data subset, so that the neural network model can dynamically adapt to changes in the spatial dimension and thus be trained to adapt to new tasks in the spatial dimension.

[0076] The above-mentioned geographic parameter values ​​refer to geography-related parameter values, such as altitude, preset area (such as power transmission line, volcano, sea or river), geographic latitude, geographic coordinates and the like.

[0077] In some embodiments, the spring data of the same year in the historical fire data set is divided into the old task data set according to the time dimension, and the autumn data of the same year in the historical fire data set is divided into the new task data set. Alternatively, the autumn data of the same year in the historical fire data set is divided into the old task data set according to the time dimension, and the winter data of the same year in the historical fire data set is divided into the new task data set. Alternatively, the 2020 data in the historical fire data set is divided into the old task data set according to the time dimension, and the 2021 data in the historical fire data set is divided into the new task data set, and so on. The above-mentioned old task data subset can be represented as TR o , the new task data subset can be expressed as TR n .

[0078] For example, the fuel data for a winter forest fire could be: vegetation type x1: 1 (indicating coniferous forest), vegetation height x2: 3.5 meters, combustion equivalent x3: 8.2 (some standard unit, such as MJ / m 3 , indicating high combustion equivalent), moisture content of combustibles x4: 0.12 (indicating 12% moisture content, low), ignition point x5: 425 degrees Celsius, calorific value x6: 20.5 joules / gram, heat release rate\(x7\): 320 joules / second. Transmission line data, voltage x8: 500 volts, resistance x9: 0.002 ohms, reactance x10: 2.5 (Ω, indicating medium reactance), susceptance x11: 0.8 (Ω, indicating low susceptance), conductivity x12: 0.005 (S, indicating medium conductivity)

[0079] The combustible data for a summer forest fire could be: vegetation type x1: 2 (indicating broad-leaved forest), vegetation height x2: 10 meters, combustion equivalent x3: 5.6 (some standard unit, such as MJ / m 3, indicating medium combustion equivalent), moisture content of combustibles x4: 0.5 (indicating 50% moisture content, medium), ignition point x5: 300 degrees Celsius, calorific value x6: 15.8 joules / gram, heat release rate x7: 180 joules / second. Transmission line data: voltage x8: 220 volts, resistance x9: 0.005 ohms, reactance x10: 1.8 (Ω, indicating medium reactance), susceptance x11: 0.6 (Ω, indicating medium susceptance), conductivity x12): 0.003 (S, indicating medium conductivity).

[0080] S103, continuously learning the original neural network model according to multiple training data subsets to obtain a neural network model after continuous learning;

[0081] In some embodiments, continuously learning the original neural network model according to multiple training data subsets to obtain the neural network model after continuous learning specifically includes:

[0082] The original neural network model determines the first fire hazard level output by the sample data in the first task according to the sample data in the second task data subset;

[0083] The neural network model determines, based on the sample data in the second task data subset, a second fire hazard level output by the sample data in the first task and a third fire hazard level output in the second task;

[0084] Determine a distillation loss between the fire hazard level output by the original neural network model for the sample data in the first task and the fire hazard level output by the neural network model for the sample data in the first task;

[0085] Determine the task loss between the fire hazard level output by the sample data in the second task and the true label corresponding to the sample data;

[0086] According to the total loss value composed of distillation loss and task loss, the weight parameters of the neural network model are updated.

[0087] It can be seen that in the implementation process of the above scheme, the weight parameters of the neural network model are updated by combining distillation loss and task-specific loss. Distillation loss can help the model retain and strengthen the learning results of old tasks as much as possible while learning new tasks. Task loss can draw on cross-task experience and enhance the generalization ability of unseen data and unknown situations. The design of the total loss function helps to balance the learning process of new and old tasks, making the entire learning process more stable and not prone to sacrificing the performance of old tasks due to over-optimization of new tasks, thereby effectively reducing the probability of "catastrophic forgetting" and improving the accuracy of fire hazard level prediction.

[0088] like Figure 4 , Figure 5 As shown, in some embodiments, the hidden layers of the deep neural network model use different colors to represent these parameter structures, and the parameter structures are different convolutional layers. The model allows some parameters to be shared between different tasks and distinguishes between specific parameters of old tasks and new tasks:

[0089] 1) Input layer: The input of the model consists of two parts:

[0090] a. Combustible data:

[0091] -Input features: Xveg = [x1, x2, x3, x4, x5, x6, x7], where: x1 represents vegetation type, x2 represents vegetation height, x3 represents combustion equivalent, x4 represents moisture content of combustibles, x5 represents ignition point, x6 represents calorific value, and x7 represents heat release rate.

[0092] b. Transmission line data:

[0093] -Input characteristics: Xpower = [x_8, x_9, x10, x11, x12], where: x8 represents voltage, x9 represents resistance, x10 represents reactance, x11 represents susceptance, and x12 represents conductance.

[0094] Before processing the input data, a convolution layer is used to extract features. For the combustible data part: Convolution kernel: K veg , convolution kernel size d veg . Convolutional layer output:

[0095] Z veg_conv =Conv1D(X veg ,K veg ,d veg ) Transmission line data part: Convolution kernel: K power , convolution kernel size d power . Convolutional layer output: Z power_conv =Conv1D(X power ,K power ,d power )

[0096] 2) Fully connected layer and joint layer structure, fully connected layer:

[0097] a. Combustible data section:

[0098] -Fully connected layer parameters: (Shared Parameters)

[0099] -Old task parameters: (Old task-specific parameters)

[0100] - New mission parameters: (new task specific parameters)

[0101] Fully connected layer output:

[0102]

[0103] b. Transmission line data section:

[0104] -Fully connected layer parameters: (Shared Parameters)

[0105] -Old task parameters: (Old task-specific parameters)

[0106] - New mission parameters: (new task specific parameters)

[0107] Fully connected layer output:

[0108]

[0109] 3) Joint layer (optional):

[0110] If there is a joint layer, the features of the two parts of data are combined and classified prediction is performed. The joint layer output is:

[0111] Z joint =ReLU(W joint ·[Z veg ; Z power ]+b joint )

[0112] 4) Output layer:

[0113] The output layer predicts the fire danger level and uses the softmax activation function for multi-classification. Output layer output:

[0114]

[0115] Preferably, in freezing the shared parameter θ shared and the weight parameter θ corresponding to the old task old After that, the weight parameter θ corresponding to the new task can be randomly initialized new , and then the original neural network is used to predict the sample data in the new task data subset to obtain the fire hazard level output by the sample data on the old task. The above prediction process can be expressed as Y o =SV o (X new ,θ shared ,θ old ), where Y oIt represents the fire danger level output by the original neural network on the old task after the sample data is input, SV o represents the original neural network, θ old represents the weight parameter corresponding to the old task, θ shared Represents shared parameters between different new and old tasks.

[0116] Among them, the above-mentioned original neural network and neural network model have the same network structure. The difference is that the original neural network is trained using a subset of the old task data, and its function is to predict the fire hazard level output by the new task data sample on the old task (that is, retain the output of the previous round of neural network for the new task data sample), however, the neural network model is being trained using a subset of the new task data, and its function is to predict the fire hazard level output by the new task data sample on the new task.

[0117] In some embodiments, the first fire hazard level and the second fire hazard level are determined according to the following formula:

[0118] y o =SV n (X new ,θ shared ,θ old )

[0119] y n =SV n (X new ,θ shared ,θ old )

[0120] Among them, y o is the fire hazard level output by the sample data in the first task of the original neural network model, y n is the fire hazard level output by the sample data in the second task of the neural network model, SV n is the neural network model, θ old is the weight parameter corresponding to the first task, θ shared is the shared parameter between the second task and the first task.

[0121] Preferably, a neural network model SV is used model , train on a new task data subset and predict sample data Fire hazard ratings on old and new tasks.

[0122]

[0123] Calculate the distillation loss and task loss, distillation loss:

[0124] The distillation loss between the fire hazard level output by the original neural network for the sample data on the old task and the fire hazard level output by the neural network model for the sample data on the old task is calculated by the formula; where H represents the selected knowledge distillation function, It represents the fire danger level output by the original neural network on the old task after inputting sample data. Represents the fire danger level output by the neural network model on the old task after inputting sample data.

[0125] Task-Specific Loss:

[0126] The task-specific loss between the fire hazard level output by the sample data on the new task and the true label corresponding to the sample data is calculated by the formula; represents the task loss between the fire hazard level output by the sample data on the new task and the true label corresponding to the sample data, L represents the selected classification loss function, such as MSE, etc. It represents the fire danger level output by the neural network model on the new task after the sample data is input, y true Indicates the true label corresponding to the sample data.

[0127] In some embodiments, according to the total loss value composed of the distillation loss and the task loss, the weight parameters of the neural network model are updated, specifically including:

[0128]

[0129] in, are the shared parameters of the neural network model between the second task and the first task, is the weight parameter of the neural network model corresponding to the first task, is the weight parameter of the neural network model corresponding to the second task, λ o To set the hyperparameters of the balanced distillation loss and task loss, is the distillation loss, is the task loss between the fire hazard level output by the sample data on the second task and the true label corresponding to the sample data, is the regularization term, θ s To update the shared parameters of the neural network model between different second tasks and the first task, θ o is the weight parameter of the neural network model corresponding to the first task before updating, θ nIt is the weight parameter corresponding to the second task of the neural network model before updating.

[0130] Preferably, in the training process of learning with some sample data in the historical fire data set, the neural network model can use the gradient descent method to continuously correct the weights and thresholds of the neural network model. The error calculation formula between the hidden layer and the output layer of the above neural network model is:

[0131] δ k =(y real -y k )y k (1-y k );

[0132] Among them, δ k Represents the error between the hidden layer and the output layer of the neural network model, y real Represents the true label (i.e., true value) corresponding to the sample data, y k represents the fire hazard level output by the neural network model for the sample label. The modified weight of the above neural network model can be expressed as w′ jk =w jk -η1×δ k ×h j , the threshold value of the modified neural network model can be expressed as q′ k =q k -η2×δ k ; Among them, w jk and w′ jk Represent the weights before and after correction respectively; q k and q′ k represent the thresholds before and after correction respectively; η1 and η2 represent the correction coefficients of weight and threshold respectively.

[0133] S104. Determine the fire hazard levels corresponding to the first task data subset and the second task data subset according to the neural network model after continuous learning.

[0134] In some embodiments, determining the fire hazard level corresponding to the first task data subset and the second task data subset according to the neural network model after continuous learning specifically includes:

[0135] The fire hazard levels corresponding to the first task data subset and the second task data subset are evaluated based on the comparison relationship between the fire occurrence probability output by the neural network model and the preset threshold. The fire hazard levels include four levels: the first hazard level, the second hazard level, the third hazard level and the fourth hazard level.

[0136] In some embodiments, the fire hazard level corresponding to the first task data subset and the second task data subset is evaluated according to the comparison relationship between the fire occurrence probability output by the neural network model and the preset threshold, specifically including:

[0137] When the probability of fire occurrence output by the neural network model is in the range of 0.9-1, the fire hazard level is the first hazard level;

[0138] When the probability of fire occurrence output by the neural network model is in the range of 0.7-0.9, the fire danger level is the second danger level;

[0139] When the probability of fire occurrence output by the neural network model is in the range of 0.6-0.7, the fire hazard level is the third hazard level;

[0140] When the probability of fire occurrence output by the neural network model is in the range of 0-0.6, the fire hazard level is the fourth hazard level.

[0141] After the continuous learning is completed, the obtained neural network model needs to be evaluated, especially the performance evaluation in predicting the fire danger level. The evaluation indicators can include:

[0142] Accuracy: The classification accuracy of the model on the test set when predicting the fire danger level (extremely high, high, medium, low).

[0143] Anti-forgetting ability: evaluates the accuracy of the model when switching between different tasks or continuously learning.

[0144] The present application also proposes a fire hazard level assessment system, the system comprising: a data acquisition unit, a data classification unit, a model training unit and an assessment unit;

[0145] A data acquisition unit, used to acquire a historical fire data set, the historical fire data set including combustible material data, power transmission line data and corresponding fire hazard levels;

[0146] A data classification unit, used for dividing the historical fire data set into a plurality of training data subsets, wherein the plurality of training subsets include a first task data subset and a second task data subset;

[0147] A model training unit, used for continuously learning the original neural network model and the neural network model being learned on the second task according to multiple training data subsets to obtain a neural network model after continuous learning;

[0148] An evaluation unit is used to determine the fire hazard level corresponding to the second task data subset according to the neural network model after continuous learning.

[0149] The present application also proposes a readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the following steps:

[0150] Obtain a historical fire data set, which includes combustible data, power transmission line data, and corresponding fire hazard levels;

[0151] Dividing the historical fire data set into a plurality of training data subsets, the plurality of training subsets comprising a first task data subset and a second task data subset, the first task data subset and the second task data subset respectively using a first structural parameter, a first training strategy and a second structural parameter, a second training strategy to train the original neural network model;

[0152] Continuously learning the original neural network model according to multiple training data subsets to obtain a neural network model after continuous learning;

[0153] The fire hazard levels corresponding to the first task data subset and the second task data subset are determined according to the neural network model after continuous learning.

[0154] The present application also proposes a computer device, including a memory and a processor, wherein the memory stores a computer program, and the computer program is executed by the processor to perform the following steps:

[0155] Obtain a historical fire data set, which includes combustible data, power transmission line data, and corresponding fire hazard levels;

[0156] Dividing the historical fire data set into a plurality of training data subsets, the plurality of training subsets comprising a first task data subset and a second task data subset, the first task data subset and the second task data subset respectively using a first structural parameter, a first training strategy and a second structural parameter, a second training strategy to train the original neural network model;

[0157] Continuously learning the original neural network model according to multiple training data subsets to obtain a neural network model after continuous learning;

[0158] The fire hazard levels corresponding to the first task data subset and the second task data subset are determined according to the neural network model after continuous learning.

[0159] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiments can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in the present application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0160] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0161] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the patent of the present application. It should be pointed out that, for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. The above disclosure is only a preferred embodiment of the present invention, and of course it cannot be used to limit the scope of the rights of the present invention. Therefore, equivalent changes made according to the claims of the present invention are still within the scope covered by the present invention.

Claims

1. A method for evaluating fire hazard level, characterized in that: The method comprises the following steps: Acquire a historical fire data set, wherein the historical fire data set includes combustible material data, power transmission line data, and corresponding fire hazard levels; Dividing the historical fire data set into a plurality of training data subsets, the plurality of training subsets comprising a first task data subset and a second task data subset, the first task data subset and the second task data subset respectively using a first structural parameter, a first training strategy and a second structural parameter, a second training strategy to train an original neural network model; Continuously learning the original neural network model according to the multiple training data subsets to obtain a neural network model after continuous learning; The fire hazard levels corresponding to the first task data subset and the second task data subset are determined according to the neural network model after continuous learning.

2. The fire hazard level assessment method according to claim 1, characterized in that: The dividing the historical fire data set into a plurality of training data subsets specifically includes: The historical fire data set is divided according to a time dimension or a space dimension to obtain a plurality of training data subsets.

3. The fire risk level assessment method according to claim 2, characterized in that: The continuously learning the original neural network model according to the multiple training data subsets to obtain the neural network model after continuous learning specifically includes: The original neural network model determines, based on the sample data in the second task data subset, a first fire hazard level output by the sample data in the first task; The neural network model determines, based on the sample data in the second task data subset, a second fire hazard level output by the sample data in the first task and a third fire hazard level output in the second task; Determine a distillation loss between the fire hazard level output by the original neural network model for the sample data in the first task and the fire hazard level output by the neural network model for the sample data in the first task; Determining a task loss between a fire hazard level output by the sample data in a second task and a true label corresponding to the sample data; According to the total loss value composed of the distillation loss and the task loss, the weight parameters of the neural network model are updated.

4. The fire risk level assessment method according to claim 3, characterized in that: The first fire hazard level and the second fire hazard level are determined according to the following formula: y o =SV n (X new ,i shared ,i old ) y n =SV n (X new ,i shared ,i old ) Among them, y o is the fire hazard level output by the sample data in the first task of the original neural network model, y n is the fire hazard level output by the sample data in the second task of the neural network model, SV n is the neural network model, θ old is the weight parameter corresponding to the first task, θ shared is the shared parameter between the second task and the first task.

5. The fire risk level assessment method according to claim 4, characterized in that: The total loss value formed by the distillation loss and the task loss and updating the weight parameters of the neural network model specifically include: in, are the shared parameters of the neural network model between the second task and the first task, is the weight parameter of the neural network model corresponding to the first task, is the weight parameter of the neural network model corresponding to the second task, λ o To set the hyperparameters of the balanced distillation loss and task loss, is the distillation loss, is the task loss between the fire hazard level output by the sample data on the second task and the true label corresponding to the sample data, is the regularization term, θ s To update the shared parameters of the neural network model between different second tasks and the first task, θ o is the weight parameter of the neural network model corresponding to the first task before updating, θ n It is the weight parameter corresponding to the second task of the neural network model before updating.

6. The fire risk level assessment method according to claim 5, characterized in that: The determining, according to the neural network model after the continuous learning, the fire hazard levels corresponding to the first task data subset and the second task data subset specifically includes: The fire hazard levels corresponding to the first task data subset and the second task data subset are evaluated based on the comparison relationship between the fire occurrence probability output by the neural network model and the preset threshold, and the fire hazard levels include four levels: a first hazard level, a second hazard level, a third hazard level and a fourth hazard level.

7. The fire risk level assessment method according to claim 6, characterized in that: The evaluating the fire hazard level corresponding to the first task data subset and the second task data subset according to the comparison relationship between the fire occurrence probability output by the neural network model and the preset threshold value specifically includes: When the fire occurrence probability output by the neural network model is in the range of 0.9-1, the fire hazard level is the first hazard level; When the fire occurrence probability output by the neural network model is in the range of 0.7-0.9, the fire hazard level is the second hazard level; When the fire occurrence probability output by the neural network model is in the range of 0.6-0.7, the fire hazard level is the third hazard level; When the fire occurrence probability output by the neural network model is in the range of 0-0.6, the fire hazard level is the fourth hazard level.

8. A fire hazard level assessment system, characterized in that: The system comprises: a data acquisition unit, a data classification unit, a model training unit and an evaluation unit; The data acquisition unit is used to acquire a historical fire data set, wherein the historical fire data set includes combustible material data, power transmission line data and corresponding fire hazard levels; The data classification unit is used to divide the historical fire data set into a plurality of training data subsets, wherein the plurality of training subsets include a first task data subset and a second task data subset; The model training unit is used to continuously learn the original neural network model and the neural network model being learned on the second task according to the multiple training data subsets to obtain the neural network model after continuous learning; The evaluation unit is used to determine the fire hazard level corresponding to the second task data subset according to the neural network model after continuous learning.

9. A readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor is caused to execute the steps of the method according to any one of claims 1 to 7.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method according to any one of claims 1 to 7.

Citation Information

Cited By

  • Intelligent agricultural four-condition comprehensive monitoring system and method

    CN119984396A